Results - GeoMLA

Statistical modeling of phenological phases
in Poland based on coupling satellite
derived products and gridded
meteorological data
Bartosz Czernecki [1]. Jakub Nowosad [1], Katarzyna Jabłońska [2]
[1]
Faculty of Geosciences
Adam Mickiewicz University in Poznań, Poland
nwp@amu.edu.pl
[2]
Institute of Meteorology and Water Management
- National Research Instiute, Warsaw, Poland
Phenological observation in
Poland – after 1945
1951-1992
1951-1992 – network of phenological
observations run by Polish Met Service (IMGW);
strongly varying number of stations (over 700 in
1970s and below 100 in 1980s). Only ~30
stations with complete and reliable dataset.
1993-2004 – network of phenological
observations completely abandonded
2005-2007 – reactivation of phenological
observation, partly in old locations (~40 stations)
2007 – onwards – newly established network in
location of meteorological stations (~60
stations). Measurements according to BBCHscale
2007 onwards
Why to use phenological data?
Global warming determine the advance of
phenological events. Therefore, changes in timing of
phenological phases are important proxies in
contemporary climate research:
- climate change
- climate proxy
- dynamics of climate seasonality
- food production
- aerobiology (pollen)
Year
Czernecki and Jabłońska
(2016) (in print)
Aim
•
The main aim → create and evaluate different statistical models for
reconstructing and predicting day of year of selected phenological phases
occurrence using the most recent data → finding a robust predictors
•
Evaluate possibilities of using only free of charge data remote sensing and
meteorological data as predictors
– (1) distinguish the amount of information provided by both sources of
data
– (2) define whether they are unrelated and contain possible sources of
not overlapping information,
– (3) and thus may (or may not) robustly contribute in phenological
research, especially in terms of phenological modeling
•
Tools → everything written in R programming languages and its packages to
automatize entire procedure
Opadlisci brzoza
Opadlisci lipa
Zolkn brzoza
Zolkn lipa
Zolkn kasztanowiec
Dojrzewanie kasztanowiec
Dojrzewanie leszczyna
Kwitnienie wrzos
Zakwitanie lipa
Zakwitanie akacja
Zakwitanie lilak
Zakwitanie kasztanowiec
Zakwitanie czeremcha
Zakwitanie mniszek
Listnienie brzoza
Zakwitanie podbiał
Zakwitanie leszczyna
0
100
Day of year
200
300
The study period covers years 2007-2014 and contains only quality-controlled
dataset of Syringa vulgaris and Aesculus hippocastanum flowering dates (i.e. late
spring phenophase) on 52 stations in Poland
Syringa vulgaris
By Ulf Eliasson - Own work, CC BY 2.5,
https://commons.wikimedia.org/w/index.php?curid=1387269
Aesculus hippocastanum
By H. Zell - Own work, CC BY-SA 3.0,
https://commons.wikimedia.org/w/index.php?curid=8951128
Error detection
(data visualisation → expert decision)
Predictor variables and
data preparation
Three types of data sources were used as predictors:
1.
2.
Satellite derived products → MODIS level-3 vegetation products:
–
NDVI, EVI, LAI, fPAR (fraction of photosynthetically active radiation)
–
Interactive Multisensor Snow and Ice Mapping System (IMS) products
–
Highly noisy data → pixel reliability information taken into account
Preprocessed gridded meteorological data → ECA&D
–
3.
cumulative growing degree days (GDD), cumulative growing precipitation
days (GPD), average monthly temperatures, monthly temperatures over
the previous year
Spatial features (longitude, latitude, altitude, distance to Baltic Sea, etc.)
Development of statistical
models
A few methods were tested and evaluated against the onset dates of phenophases:
–
multiple linear regression with (lmAIC) and without stepwise selection (lm)
–
generalized linear model with (glmAIC) and without stepwise selection
(glm)
–
random forest (RF)
Potential predictors splited into four sub-groups (to estimate their importance):
1.
Only meteorological-derived variables and locations’ features
2.
MODIS-derived predictors
3.
All available variables pre-processed with the use of Boruta algorithm (that
finds all-relevant features) (Kursa 2010)
4.
All available variables without any pre-selection
Cross-validation
Repeated k-fold cross validation was used to avoid overfitting and to estimate
the accuracy of the models
„Built-in” k-fold approach in caret package led to overfiting of models → spliting
data on annual basis to avoid overfiting
2007 2008 2009 2010 2011 ...
Results
Substantial impact of
predictors selection on final
results
●
●
The chosen set of predictors →
very similar results in every of
tested regression based
algorithms
The impact of selected
predictors was smaller on RF
than on regression models.
However, the obtained pattern
was similar in every of
analyzed models, i.e. the best
fitted models were
preprocessed using Boruta
algorithm.
Results
Substantial impact of
predictors selection on
final results:
●
The AIC stepwise
screening hardly influences
the obtained results → do
not redress computational
time that is required while
applying this procedure
Results
●
●
Models based only on meteorological indices accounted for about 80% of
variance in Syringa vulgaris and Aesculus hippocastanum flowering dates,
applying remote sensing data and preprocessing by Boruta algorithm
increased this value to over 95%
Conclusions
●
●
●
●
●
●
The created models show high applicable potential
The models based on meteorological characteristics were better fitted to
observational time-series than remote sensing-based models
Even though, conjunction of both data sources improve model’s accuracy
A strong improvement if preprocessing procedures (e.g. Boruta) were
applied → numerous set of potential predictors
Clear limitations of applying satellite observation in phenology modeling:
–
small contribution of satellite derived products to model’s results
–
satellite data contain noisy information and thus were omitted while
applying preprocessing procedures
Therefore, most of the created phenology models are primarily based on
climatological indices with only slight improvements of satellite products
Thank you for your attention